Methods, systems, and equipment for super-resolution remote sensing image reconstruction based on frequency domain enhancement

Abstract

This invention belongs to the field of image data processing technology, specifically relating to a method, system, and device for super-resolution remote sensing image reconstruction based on frequency domain enhancement. The method includes: S1, extracting shallow features from a low-resolution remote sensing image; S2, inputting the shallow features into multiple cascaded frequency interaction processing steps, performing bi-branch processing on the input features, obtaining a first feature by radially weighting and inverse transforming different frequency components in the frequency domain, and obtaining a second feature through depthwise separable convolution, activation function, selection of scanning modules, and layer normalization; fusing the two branches of features according to weights and then skip-connecting them with the original features, followed by feedforward network enhancement, repeating this process a set number of times, and then convolution and residual connection to obtain deep features; S3, fusing the deep and shallow features, performing convolution and pixel rearrangement upsampling, and outputting a super-resolution image. This invention, while maintaining structural consistency, can more accurately reconstruct the texture details and edge contours of remote sensing images.

Description

Technical Field

[0001] This invention belongs to the field of image data processing technology, specifically relating to a method, system, and device for super-resolution remote sensing image reconstruction based on frequency domain enhancement. Background Technology

[0002] Remote sensing technology plays a vital role in fields such as geological exploration, ecological environment monitoring, and disaster early warning. However, due to limitations in sensor physical performance, imaging conditions, and transmission and storage, the actual acquired remote sensing images often suffer from low spatial resolution, making it difficult to meet the needs of refined ground feature identification and quantitative analysis.

[0003] Traditional super-resolution reconstruction methods based on interpolation or frequency domain signal processing are simple to implement and computationally efficient, but they perform poorly in texture reconstruction and edge preservation when dealing with complex terrain structures, and are prone to introducing blurring effects or producing unnatural artifacts.

[0004] Due to the wide coverage, highly complex land cover types, significant scale differences, and strong diversity of texture structures in remote sensing images, existing deep learning-based methods still face severe challenges in processing highly heterogeneous surface areas. For example, in complex scenes such as building outlines, the subtle textures of dense vegetation, and the reflective features of water surfaces, existing models are prone to problems such as texture blurring, edge distortion, or structural inconsistencies, resulting in defects in the naturalness and structural realism of the generated images. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and device for super-resolution remote sensing image reconstruction based on frequency domain enhancement.

[0006] Super-resolution remote sensing image reconstruction methods based on frequency domain enhancement include:

[0007] S1. Obtain low-resolution remote sensing images and perform shallow feature extraction to obtain shallow features;

[0008] S2. Shallow features are processed by a series of sequentially connected frequencies to obtain deep features. Each frequency interaction is processed as follows:

[0009] S2.1. Perform the following branching processing on shallow features simultaneously:

[0010] Branch 1: The shallow features are subjected to channel dilation and then frequency domain transformation to obtain a frequency domain map. In the frequency domain map, weights are applied to different frequency domain components according to the center position of the frequency domain map to obtain a weighted frequency domain map. The first branch features are obtained by inverse Fourier transform and the first channel weights are obtained by processing through the channel selection interaction mechanism.

[0011] Branch 2: The shallow features are processed sequentially through depth-separable convolution, activation, selective scanning, and layer normalization to obtain the second branch features. The second channel weights are obtained through a channel selection interaction mechanism.

[0012] S2.2. The first branch features and the second branch features are weighted and fused by Hadamard product according to the first channel weight and the second channel weight to obtain the fused features. The fused features are then connected with the shallow features to obtain the first features. The first features are then enhanced to obtain the second features. The first features and the second features are then connected to obtain the third features.

[0013] S2.3. Use the third feature as a new shallow feature and execute S2.1. Then use the new third feature as a new shallow feature and execute S2.1 again. Repeat this process for a set number of times. Perform convolution and residual connection on all the results to obtain the fourth feature. Then perform a skip connection between the fourth feature and the shallow feature to obtain the deep feature.

[0014] S3. Deep features and shallow features are skipped and connected, then convolutional and pixel rearrangement are performed, followed by upsampling to obtain a super-resolution image.

[0015] The specific steps of the selection scanning process in branch two of S2.1 are as follows: the feature map after activation is linearly mapped into four channel groups. Each channel group flattens the two-dimensional image features into a one-dimensional sequence according to a different scanning strategy. After structured state space sequence processing, a spatial model representation of the continuous time state is obtained. The spatial model representation is converted into a discrete model and convolved in a recursive form to obtain the input for layer normalization processing.

[0016] The scanning strategy is a row-first and column-first scanning method. The four scanning methods are row-first forward scanning, row-first reverse scanning, column-first forward scanning, and column-first reverse scanning.

[0017] Convolution is performed recursively, specifically as follows:

[0018] ,

[0019] ,

[0020] in, It is a structured convolution kernel, a matrix , , and These are the discrete-time state transition matrix, the discrete-time state input matrix, the output matrix, and the direct transfer matrix, respectively. It is the length of the input sequence. Given the input sequence, This is the output sequence, which is the input for the normalization process in the middle layer of branch two.

[0021] In S2.2, the feature enhancement of the first feature is as follows: the first feature is normalized, and after normalization, the channel dimension is expanded by linear projection. Then, spatial modeling and local information extraction are performed by depthwise separable convolution. After processing, activation is performed to obtain the second feature.

[0022] After processing, activation is performed to obtain the second feature, which includes Gaussian error linear unit activation and Sigmoid weighted linear activation.

[0023] In S2.1, branch one applies weights to different frequency components based on the center position of the frequency domain graph, specifically as follows:

[0024] ,

[0025] in, To assign weights to different frequency domain components, H and W represent the height and width of the frequency domain plot, respectively. Represents the position coordinates in the frequency domain plot. This represents the coordinates of the center position of the frequency domain plot.

[0026] In S2.1, after the first branch feature channel selection interaction mechanism in branch one is processed, global average pooling is then performed to obtain the weight of the first channel.

[0027] A frequency-domain-enhanced super-resolution remote sensing image reconstruction system, used to implement the aforementioned frequency-domain-enhanced super-resolution remote sensing image reconstruction method, includes:

[0028] The shallow feature acquisition module acquires low-resolution remote sensing images and extracts shallow features to obtain shallow features.

[0029] The deep feature acquisition module obtains deep features by processing shallow features through a series of sequentially connected frequency interactions. Each frequency interaction is processed as follows:

[0030] The following branching processes are performed simultaneously on the shallow features:

[0031] Branch 1: The shallow features are subjected to channel dilation and then frequency domain transformation to obtain a frequency domain map. In the frequency domain map, weights are applied to different frequency domain components according to the center position of the frequency domain map to obtain a weighted frequency domain map. The first branch features are obtained by inverse Fourier transform and the first channel weights are obtained by processing through the channel selection interaction mechanism.

[0032] Branch 2: The shallow features are processed sequentially through depth-separable convolution, activation, selective scanning, and layer normalization to obtain the second branch features. The second channel weights are obtained through a channel selection interaction mechanism.

[0033] The first branch features and the second branch features are weighted and fused together by Hadamard product according to the first channel weight and the second channel weight to obtain the fused feature. The fused feature is then connected with the shallow features to obtain the first feature. The first feature is then enhanced to obtain the second feature. The first feature and the second feature are then connected to obtain the third feature.

[0034] The third feature is used as a new shallow feature to perform branch processing. The resulting new third feature is used as a new shallow feature to perform branch processing again. This process is repeated a set number of times. All the results are then convolved and residual connected to obtain the fourth feature. The fourth feature is then skipped and connected with the shallow features to obtain the deep feature.

[0035] The super-resolution image output module performs skip connections between deep and shallow features, convolution and pixel rearrangement, and upsampling to obtain a super-resolution image.

[0036] A frequency-domain-enhanced super-resolution remote sensing image reconstruction device includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the frequency-domain-enhanced super-resolution remote sensing image reconstruction method.

[0037] This invention provides a super-resolution remote sensing image reconstruction method based on frequency domain enhancement, which has the following beneficial effects:

[0038] Branch one of this invention introduces frequency domain enhancement processing, employing a radial frequency weighting strategy in the frequency domain to effectively strengthen high-frequency components related to texture and edges, while suppressing low-frequency redundant information. This mechanism enables more accurate recovery of detailed features when dealing with complex remote sensing scenes such as building outlines, fine textures of dense vegetation, and water surface reflections, significantly mitigating the texture blurring and edge distortion problems commonly found in traditional methods.

[0039] To address the challenges of complex land cover types, significant scale differences, and high structural diversity in remote sensing images, this invention designs a branching processing mechanism to achieve multi-dimensional feature interaction, enabling selective feature fusion across channel and sequence dimensions. Combined with the long-range dependency modeling capabilities of selective scan processing, this application can better understand and represent highly heterogeneous surface structures, improving the structural consistency and naturalness of reconstructed images.

[0040] Furthermore, the lightweight depthwise separable convolution processing and efficient selective scanning processing control computational complexity while ensuring strong feature extraction and reconstruction capabilities. Compared to traditional interpolation- or frequency domain-based methods, this method significantly improves visual quality and structural fidelity; compared to some complex deep learning models, this method achieves a better balance between inference efficiency and detail recovery.

[0041] Meanwhile, this application employs a smooth nonlinear activation function to further enhance the expressive power of features. Through the above design, this invention can more accurately reconstruct the texture details and edge contours of remote sensing images while maintaining structural consistency. Attached Figure Description

[0042] Figure 1 shows a low-resolution remote sensing image;

[0043] Figure 2 The image is a super-resolution remote sensing image obtained by the method described in this application. Detailed Implementation

[0044] Example

[0045] To further understand the content of this invention, the invention will be described in detail with reference to the embodiments.

[0046] This invention relates to a method, system, and device for super-resolution remote sensing image reconstruction based on frequency domain enhancement, comprising the following steps:

[0047] Super-resolution remote sensing image reconstruction methods based on frequency domain enhancement include:

[0048] S1. Obtain low-resolution remote sensing images and perform shallow feature extraction to obtain shallow features.

[0049] For an input low-resolution image Shallow features are obtained by passing through a 3×3 convolutional layer. This process can be represented as:

[0050] ,

[0051] in, It is a shallow feature obtained after a 3×3 convolution layer. H, W, and C represent the height, width, and number of channels of the feature map, respectively. The number of channels is determined by the number of convolution kernels. This indicates a convolutional layer with a kernel size of 3x3. This represents a low-resolution image with a height of H, a width of W, and 3 channels.

[0052] The shallow feature extraction part captures the basic edge, texture and brightness information of the remote sensing image, providing the original foundation for subsequent deep feature extraction.

[0053] S2. Shallow features are processed by a series of sequentially connected frequencies to obtain deep features. Each frequency interaction is processed as follows:

[0054] S2.1. Perform the following branching processing on shallow features simultaneously:

[0055] Branch 1: The shallow features are subjected to channel dilation and then frequency domain transformation to obtain a frequency domain map. In the frequency domain map, weights are applied to different frequency domain components according to the center position of the frequency domain map to obtain a weighted frequency domain map. The first branch features are obtained by inverse Fourier transform and the first channel weights are obtained by processing through the channel selection interaction mechanism.

[0056] Branch 2 involves sequentially processing the shallow features through depthwise separable convolution, activation, selective scanning, and layer normalization to obtain the second branch features. The second channel weights are then obtained through a channel selection interaction mechanism.

[0057] Specifically, one-third of the input features are dilated into channels by linear projection. C, This is a predefined channel expansion factor, which is set to 1 here.

[0058] During frequency domain enhancement, the input features are transformed in the frequency domain, projecting the spatial domain features onto a two-dimensional frequency domain map. :

[0059] ,

[0060] in, Represents the Fast Fourier Transform. This represents a linear projection operation, where U is the input feature of branch one, which is a shallow feature here.

[0061] In the frequency domain image, weights are applied to different frequency components based on their center position to obtain a weighted frequency domain image. This highlights high-frequency information related to texture and edges while suppressing low-frequency redundant components. (Weighted frequency domain image) for: ,

[0062] ,

[0063] in, To assign weights to different frequency domain components, H and W represent the height and width of the frequency domain plot, respectively. Represents the position coordinates in the frequency domain plot. This represents the coordinates of the center position of the frequency domain plot. This represents the weighting factor.

[0064] Weighted frequency domain graph via inverse Fourier transform Returning to the spatial domain, we obtain the features of the first branch:

[0065] ,

[0066] The weight of the first channel is obtained through a channel selection interaction mechanism. Specifically, after the first branch feature channel selection interaction mechanism is processed, global average pooling is then performed to obtain the weight of the first channel.

[0067] ,

[0068] GAP(⋅) represents the Global Average Pooling operation, used to aggregate global information in the spatial dimension; GELU(⋅) is the Gaussian Error Linear Unit activation function, used to enhance the nonlinear expressive power of the model; This indicates a convolution operation with a kernel size of 1×3; Sigmoid(⋅) represents the Sigmoid activation function, used to normalize the input features. Branch two also requires processing the input features. Expand the channel using linear projection. C, here the input features still refer to shallow features.

[0069] The second branch features are obtained by sequentially performing depthwise separable convolution, activation processing, selective scanning processing, and layer normalization. The process can be described as follows:

[0070] ,

[0071] Among them, select scan processing Specifically, the activated feature map is linearly mapped into four channel groups. Each channel group flattens the two-dimensional image features into a one-dimensional sequence according to a different scanning strategy. After structured state space sequence processing, a spatial model representation of the continuous time state is obtained, which can effectively capture long-distance dependencies in the sequence. Through recursive updating of the hidden state vector, efficient modeling and information transmission of the sequence are achieved, thereby improving the ability to express and predict complex time series data. The spatial model representation is converted into a discrete model and recursively convolved to obtain the input for layer normalization processing.

[0072] The scanning strategy is a row-first and column-first scanning method. The four scanning methods are row-first forward scanning, row-first reverse scanning, column-first forward scanning, and column-first reverse scanning.

[0073] The spatial model of continuous-time state is represented as:

[0074] ,

[0075] in, It is the derivative of the hidden state vector with respect to time t; It is a hidden state vector, representing the internal state of the system at time t, and a matrix. It is a continuous-time state transition matrix that determines the dynamic changes of the hidden state vector; the matrix It is a continuous-time state input matrix, representing how the input signal affects the change of the hidden state vector; It is the input signal. It is the output signal; matrix It is the output matrix, representing how the hidden state vectors affect the output; scalar It is a direct transfer matrix, representing the direct effect of input on output.

[0076] Convert a continuous model to a discrete model:

[0077] ,

[0078] ,

[0079] in, It is the time step, representing the sampling interval. It is a discrete-time state transition matrix. For matrix exponents, It is a discrete-time state input matrix. It is the identity matrix, a matrix It is a continuous-time state transition matrix. It is a continuous-time state input matrix;

[0080] After discretizing the spatial model, the recurrence equation is obtained:

[0081] ;

[0082] in, It is the first The hidden state vector at each time step It is the first Input at each time step, It is the first The output of each time step It is the length of the input sequence, the matrix , , and These are the discrete-time state transition matrix, discrete-time state input matrix, output matrix, and direct transfer matrix, which control state updates and output generation, respectively.

[0083] Convolution is performed recursively to facilitate parallel computation, specifically as follows:

[0084] ,

[0085] ,

[0086] in, It is a structured convolution kernel, a matrix , , and These are the discrete-time state transition matrix, the discrete-time state input matrix, the output matrix, and the direct transfer matrix, respectively. It is the length of the input sequence. Given the input sequence, The output sequence serves as the input for the normalization process in the second branch. Finally, all sequences are merged and reshaped to restore the 2D structure.

[0087] Second branch features Sequence weights are obtained through a sequence selection interaction mechanism. :

[0088] .

[0089] S2.2. The first branch features and the second branch features are weighted and fused by Hadamard product according to the first channel weight and the second channel weight to obtain the fused features. The fused features are then connected with the shallow features to obtain the first features. The first features are then enhanced to obtain the second features. The first features and the second features are then connected to obtain the third features.

[0090] Specifically, using Hadamah (a type of calculus) The weights obtained from the two branches , and characteristics , The features are then fused, and the fused features are concatenated. Finally, a linear projection is used to project the concatenated feature channels onto the input channel number. C , obtain fusion features To ensure that its shape matches the input shape:

[0091] ,

[0092] in This indicates a feature splicing operation.

[0093] The first feature is obtained by skipping connections between the fused feature and the shallow feature. The first feature is normalized, and the channel dimension is expanded by linear projection after normalization. Then, spatial modeling and local information extraction are performed by depthwise separable convolution. After processing, activation is performed to obtain the second feature.

[0094] Specifically, the first feature after normalization Z First, the channel dimension is expanded to [a certain value] using linear projection. CSubsequently, depthwise separable convolution is used to perform spatial modeling and local information extraction on the features, resulting in the feature... Y :

[0095] ,

[0096] After processing, activation is performed to obtain the second feature, which includes Gaussian error linear unit activation and Sigmoid weighted linear activation. Y To enhance feature representation, Gaussian error linear units (GELU) and sigmoid weighted linear units (SiLU) are activated respectively. The activated features are then fused, and the channel number is linearly projected back. C The second feature is obtained, and this process is represented as follows:

[0097] .

[0098] By skipping the first and second features, we obtain the third feature.

[0099] S2.3. Using the third feature as a new shallow feature, execute S2.1 again. Repeat this process, executing S2.1 a set number of times. Perform convolution and residual connections on all the results to obtain the fourth feature. Then, perform a skip connection between the fourth feature and the shallow features to obtain the deep feature. .

[0100] The ability to model high-order semantics and high-frequency textures is improved by using residual connections and convolution processing in the deep feature extraction part.

[0101] S3, Deep Features shallow features Skip connections, perform convolution and pixel rearrangement processing, and upsampling to obtain a super-resolution image.

[0102] This step includes two parts. The algorithm employs a convolutional layer and a pixel rearrangement process (PS). The PS layer rearranges the channel dimensions of the feature map, mapping the channel dimension information to the spatial dimension, achieving efficient upsampling without interpolation loss to recover high-resolution images. :

[0103] ,

[0104] in, Indicates magnification factor. , These represent the height and width of the reconstructed high-resolution image, respectively; both are the same size as the input low-resolution image. times.

[0105] A frequency-domain-enhanced super-resolution remote sensing image reconstruction system, used to implement the aforementioned frequency-domain-enhanced super-resolution remote sensing image reconstruction method, includes:

[0106] The shallow feature acquisition module acquires low-resolution remote sensing images and extracts shallow features to obtain shallow features.

[0107] The deep feature acquisition module obtains deep features by processing shallow features through a series of sequentially connected frequency interactions. Each frequency interaction is processed as follows:

[0108] The following branching processes are performed simultaneously on the shallow features:

[0109] Branch 1: The shallow features are subjected to channel dilation and then frequency domain transformation to obtain a frequency domain map. In the frequency domain map, weights are applied to different frequency domain components according to the center position of the frequency domain map to obtain a weighted frequency domain map. The first branch features are obtained by inverse Fourier transform and the first channel weights are obtained by processing through the channel selection interaction mechanism.

[0110] Branch 2: The shallow features are processed sequentially through depth-separable convolution, activation, selective scanning, and layer normalization to obtain the second branch features. The second channel weights are obtained through a channel selection interaction mechanism.

[0111] The first branch features and the second branch features are weighted and fused together by Hadamard product according to the first channel weight and the second channel weight to obtain the fused feature. The fused feature is then connected with the shallow features to obtain the first feature. The first feature is then enhanced to obtain the second feature. The first feature and the second feature are then connected to obtain the third feature.

[0112] The third feature is used as a new shallow feature to perform branch processing. The resulting new third feature is used as a new shallow feature to perform branch processing again. This process is repeated a set number of times. All the results are then convolved and residual connected to obtain the fourth feature. The fourth feature is then skipped and connected with the shallow features to obtain the deep feature.

[0113] The super-resolution image output module performs skip connections between deep and shallow features, convolution and pixel rearrangement, and upsampling to obtain a super-resolution image.

[0114] A frequency-domain-enhanced super-resolution remote sensing image reconstruction device includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the frequency-domain-enhanced super-resolution remote sensing image reconstruction method.

Claims

1. A super-resolution remote sensing image reconstruction method based on frequency domain enhancement, characterized in that, include: S1. Obtain low-resolution remote sensing images and perform shallow feature extraction to obtain shallow features; S2. Shallow features are processed by a series of sequentially connected frequencies to obtain deep features. Each frequency interaction is processed as follows: S2.

1. Perform the following branching processing on shallow features simultaneously: Branch 1: The shallow features are subjected to channel dilation and then frequency domain transformation to obtain a frequency domain map. In the frequency domain map, weights are applied to different frequency domain components according to the center position of the frequency domain map to obtain a weighted frequency domain map. The first branch features are obtained by inverse Fourier transform and the first channel weights are obtained by processing through the channel selection interaction mechanism. Branch 2: The shallow features are processed sequentially through depth-separable convolution, activation, selective scanning, and layer normalization to obtain the second branch features. The second channel weights are obtained through a channel selection interaction mechanism. The scanning process is as follows: the activated feature map is linearly mapped into four channel groups. Each channel group flattens the two-dimensional image features into a one-dimensional sequence according to a different scanning strategy. After the structured state space sequence is processed, a spatial model representation of the continuous time state is obtained. The spatial model representation is converted into a discrete model and convolved in a recursive manner to obtain the input for layer normalization. The scanning strategy is a row-first and column-first scanning method. The four scanning methods are row-first forward scanning, row-first reverse scanning, column-first forward scanning, and column-first reverse scanning. S2.

2. The first branch features and the second branch features are weighted and fused by Hadamard product according to the first channel weight and the second channel weight to obtain the fused features. The fused features are then connected with the shallow features to obtain the first features. The first features are then enhanced to obtain the second features. The first features and the second features are then connected to obtain the third features. The feature enhancement of the first feature is as follows: the first feature is normalized, and after normalization, the channel dimension is expanded by linear projection. Then, spatial modeling and local information extraction are performed by depthwise separable convolution. After processing, activation is performed to obtain the second feature. S2.

3. Use the third feature as a new shallow feature and execute S2.

1. Then use the new third feature as a new shallow feature and execute S2.1 again. Repeat this process for a set number of times. Perform convolution and residual connection on all the results to obtain the fourth feature. Then perform a skip connection between the fourth feature and the shallow feature to obtain the deep feature. S3. Deep features and shallow features are skipped and connected, then convolutional and pixel rearrangement are performed, followed by upsampling to obtain a super-resolution image.

2. The super-resolution remote sensing image reconstruction method based on frequency domain enhancement according to claim 1, characterized in that, Convolution is performed recursively, specifically as follows: ， ， in, It is a structured convolution kernel, a matrix , , and These are the discrete-time state transition matrix, the discrete-time state input matrix, the output matrix, and the direct transfer matrix, respectively. It is the length of the input sequence. Given the input sequence, This is the output sequence, which is the input for the normalization process in the middle layer of branch two.

3. The super-resolution remote sensing image reconstruction method based on frequency domain enhancement according to claim 1, characterized in that, The second feature is obtained after the processing and activation, including Gaussian error linear unit activation and Sigmoid weighted linear activation.

4. The super-resolution remote sensing image reconstruction method based on frequency domain enhancement according to claim 1, characterized in that, In S2.1, branch one applies weights to different frequency components based on the center position of the frequency domain graph, specifically as follows: ， in, To assign weights to different frequency domain components, H and W represent the height and width of the frequency domain plot, respectively. Represents the position coordinates in the frequency domain plot. This represents the coordinates of the center position of the frequency domain plot.

5. The super-resolution remote sensing image reconstruction method based on frequency domain enhancement according to claim 1, characterized in that, In S2.1, the first channel weight is obtained through the channel selection interaction mechanism. Specifically, after the first branch feature channel selection interaction mechanism is processed, global average pooling is performed to obtain the first channel weight.

6. A frequency-domain-enhanced super-resolution remote sensing image reconstruction system, used to implement the frequency-domain-enhanced super-resolution remote sensing image reconstruction method according to any one of claims 1-5, characterized in that, include: The shallow feature acquisition module acquires low-resolution remote sensing images and extracts shallow features to obtain shallow features. The deep feature acquisition module obtains deep features by processing shallow features through a series of sequentially connected frequency interactions. Each frequency interaction is processed as follows: The following branching processes are performed simultaneously on the shallow features: Branch 1: The shallow features are subjected to channel dilation and then frequency domain transformation to obtain a frequency domain map. In the frequency domain map, weights are applied to different frequency domain components according to the center position of the frequency domain map to obtain a weighted frequency domain map. The first branch features are obtained by inverse Fourier transform and the first channel weights are obtained by processing through the channel selection interaction mechanism. Branch 2: The shallow features are processed sequentially through depth-separable convolution, activation, selective scanning, and layer normalization to obtain the second branch features. The second channel weights are obtained through a channel selection interaction mechanism. The scanning process is as follows: the activated feature map is linearly mapped into four channel groups. Each channel group flattens the two-dimensional image features into a one-dimensional sequence according to a different scanning strategy. After the structured state space sequence is processed, a spatial model representation of the continuous time state is obtained. The spatial model representation is converted into a discrete model and convolved in a recursive manner to obtain the input for layer normalization. The scanning strategy is a row-first and column-first scanning method. The four scanning methods are row-first forward scanning, row-first reverse scanning, column-first forward scanning, and column-first reverse scanning. The first branch features and the second branch features are weighted and fused together by Hadamard product according to the first channel weight and the second channel weight to obtain the fused feature. The fused feature is then connected with the shallow features to obtain the first feature. The first feature is then enhanced to obtain the second feature. The first feature and the second feature are then connected to obtain the third feature. The feature enhancement of the first feature is as follows: the first feature is normalized, and after normalization, the channel dimension is expanded by linear projection. Then, spatial modeling and local information extraction are performed by depthwise separable convolution. After processing, activation is performed to obtain the second feature. The third feature is used as a new shallow feature to perform branch processing. The resulting new third feature is used as a new shallow feature to perform branch processing again. This process is repeated a set number of times. All the results are then convolved and residual connected to obtain the fourth feature. The fourth feature is then skipped and connected with the shallow features to obtain the deep feature. The super-resolution image output module performs skip connections between deep and shallow features, convolution and pixel rearrangement, and upsampling to obtain a super-resolution image.

7. A super-resolution remote sensing image reconstruction device based on frequency domain enhancement, characterized in that, It includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the frequency domain enhancement-based super-resolution remote sensing image reconstruction method as described in any one of claims 1-5.

Images